Early monitoring system for pancreatic cancer based on big data

Abstract

The present invention relates to an early monitoring system for pancreatic cancer based on big data, including a data acquisition module, a feature extraction module, a risk assessment module and an early warning trigger module. The data acquisition module is used to obtain the user's laboratory test data, imaging test data and clinical symptom data. The feature extraction module is connected to the data acquisition module, and constructs a comprehensive feature vector by extracting biochemical indicator features, imaging features and symptom features. The risk assessment module uses a pre-trained deep learning model to analyze the feature vector and generate a pancreatic cancer risk probability. The early warning trigger module compares the risk probability with a preset threshold, generates an early warning signal when the risk probability exceeds the threshold, and pushes it to a medical service terminal to suggest further diagnosis. The present invention combines multimodal data and deep learning technology to achieve accurate early monitoring of pancreatic cancer and improve diagnostic efficiency and accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to an early-stage pancreatic cancer monitoring system based on big data. Background Art

[0002] In recent years, pancreatic cancer, as a malignant tumor with a high mortality rate, faces major challenges in its early detection and diagnosis. Current medical detection methods mainly rely on manual analysis of laboratory test data (such as serum CA19-9 and CEA levels), imaging examinations (such as pancreatic structure and pancreatic duct dilatation), and clinical symptoms (such as jaundice and abdominal pain). With the development of big data and artificial intelligence technologies, some studies have attempted to use these technologies to assist in the early diagnosis of pancreatic cancer, but the existing system still focuses on a single data type and fails to fully integrate multimodal data.

[0003] Existing technologies are deficient in data integration and analysis capabilities, which are manifested in the following ways: first, there is a lack of systematic correlation analysis between data types, and the potential relationship between biochemical, imaging and symptom data cannot be fully utilized; second, the sensitivity and specificity of the risk assessment model are low, which can easily lead to false positives or false negatives; third, the triggering mechanism of the early warning system is relatively simple, and lacks precise response to individual differences and dynamic changes.

[0004] In view of the above problems, the present invention provides an early monitoring system for pancreatic cancer based on big data. Summary of the invention

[0005] The present application provides an early monitoring system for pancreatic cancer based on big data to improve the efficiency and accuracy of pancreatic cancer diagnosis.

[0006] The present application provides a pancreatic cancer early monitoring system based on big data, comprising:

[0007] A data acquisition module is used to obtain the user's laboratory test data, imaging test data and clinical symptom data, wherein the laboratory test data include serum CA19-9 level, serum CEA level and blood sugar level, the imaging test data include pancreatic parenchyma density value and pancreatic duct dilatation degree data, and the clinical symptom data include abdominal pain degree, weight change and jaundice index;

[0008] A feature extraction module, connected to the data acquisition module, for extracting feature parameters from the acquired data and constructing a feature vector, wherein the feature vector includes a biochemical index feature sub-vector, an image feature sub-vector and a symptom feature sub-vector;

[0009] A risk assessment module, connected to the feature extraction module, for obtaining the risk probability of pancreatic cancer based on the feature vector using a pre-trained deep learning model;

[0010] The early warning trigger module is connected to the risk assessment module and is used to compare the pancreatic cancer risk probability with a preset threshold, generate an early warning signal when the risk probability exceeds the preset threshold, and push the early warning signal to the medical service terminal to suggest further diagnosis.

[0011] Furthermore, the data acquisition module is specifically used for:

[0012] Time series features were generated based on the change rate of serum CA19-9 levels in laboratory examination data collected at consecutive time points and the dynamic changes in the degree of pancreatic duct dilatation in imaging examination data.

[0013] Furthermore, the feature extraction module is specifically used for:

[0014] Receive continuous time point data of serum CA19-9 level, serum CEA level and blood glucose level, and normalize the data of each indicator to eliminate the influence of different dimensions;

[0015] Based on the normalized data of multiple time points, calculate the change trend and fluctuation range of each indicator, and identify abnormal feature points that exceed the preset threshold;

[0016] The distribution pattern of the abnormal feature points is compared with historical data to generate a set of time series feature values ​​as the basis of the biochemical indicator feature sub-vector.

[0017] Furthermore, the feature extraction module is specifically used for:

[0018] Receiving data on the density value of the pancreatic parenchyma and the degree of dilatation of the pancreatic duct, and preprocessing the image data, including denoising and segmenting the target area;

[0019] Extract pancreatic edge structure and pancreatic duct morphological features from the segmented image region, including the length, diameter and curvature of the pancreatic duct;

[0020] The extracted morphological features are combined with the pancreatic density value to construct the image feature sub-vector.

[0021] Furthermore, the feature extraction module is specifically used for:

[0022] Receive the user's self-reported data on abdominal pain severity, weight change, and jaundice index, and classify the data to distinguish between acute and chronic symptoms;

[0023] For acute symptoms, assess their frequency and duration; for chronic symptoms, analyze their cumulative change trends;

[0024] The symptom severity score is compared with the user's previous health records to generate a symptom feature subvector.

[0025] Furthermore, the early warning trigger module is specifically used for:

[0026] When the risk probability is in the mild risk range, a low-priority warning signal is generated and pushed to the user terminal. It is recommended to monitor and record health data regularly;

[0027] When the risk probability is in the medium risk range, a medium priority warning signal is generated and pushed to the user's attending physician's terminal, suggesting that imaging examinations and blood tests be arranged;

[0028] When the risk probability is in a high-risk range, a high-priority warning signal is generated and pushed to the emergency notification system of the medical institution, along with the user's key health data, and it is recommended to immediately start a comprehensive diagnostic process.

[0029] Furthermore, the deep learning model includes a feature fusion network layer, a temporal feature extraction layer, and a risk prediction layer;

[0030] The feature fusion network layer is used to receive and process input features of different modalities, including a first feature processing unit, which is used to reduce the dimension and extract features of the biochemical index feature subvectors through three fully connected layers, and use a nonlinear activation function to optimize the expression of features to obtain a biochemical feature representation; a second feature processing unit, which is used to extract the spatial feature relationship of the image data through a convolution operation, and generate an image feature representation through a fully connected layer; a third feature processing unit, which is used to perform feature mapping on the symptom feature subvector to obtain a symptom feature representation; a dynamic weighted fusion unit, which is used to receive the biochemical feature representation, the image feature representation, and the symptom feature representation; and based on the received data, calculate the feature weights using an attention mechanism to generate a fused feature vector;

[0031] The temporal feature extraction layer is used to analyze the temporal evolution law of the fused feature vector, including a temporal data cache unit, which is used to maintain a set of historical fused feature vectors and construct a temporal feature matrix; a gated recurrent unit, which is used to analyze the temporal evolution law of the temporal feature matrix and generate a hidden state vector; and a multi-head attention mechanism, which is used to enhance the features of the hidden state vector and generate an enhanced temporal feature vector.

[0032] The risk prediction layer is used to generate the final risk assessment result, including a feature aggregation unit, which is used to receive the fused feature vector and the enhanced time series feature vector, realize feature merging through a skip connection structure, and generate a comprehensive feature vector using a dimensionality reduction operation; a risk probability calculation unit, which is used to perform feature mapping on the comprehensive feature vector through multiple fully connected layers, and output the pancreatic cancer risk probability distribution using a Softmax function.

[0033] This application has the following beneficial technical effects:

[0034] (1) The data collection module integrates laboratory test data, imaging test data and clinical symptom data to comprehensively capture pancreatic cancer-related features, making up for the limitations of single data source analysis, thereby significantly improving the accuracy and comprehensiveness of early risk assessment of pancreatic cancer. (2) The feature extraction module and risk assessment module are used to process multimodal feature vectors, and through the pre-trained deep learning model, accurate prediction of pancreatic cancer risk is achieved, the modeling ability of complex feature relationships is enhanced, and missed diagnosis and misdiagnosis are reduced. (3) The warning trigger module generates a graded warning signal based on the risk probability and the dynamically adjusted preset threshold, and pushes it to the medical service terminal in a timely manner, optimizing the linkage efficiency between patients and medical institutions, and providing a scientific basis for further diagnosis and intervention. (4) Through comprehensive data collection, accurate risk assessment and real-time warning functions, this system effectively improves the early detection rate of pancreatic cancer, while providing patients with personalized diagnosis and treatment recommendations, significantly improving the timeliness of treatment and prognosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a schematic diagram of a pancreatic cancer early monitoring system based on big data provided in the first embodiment of the present application. DETAILED DESCRIPTION

[0036] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0037] The first embodiment of the present application provides an early monitoring system for pancreatic cancer based on big data. Figure 1 , which is a schematic diagram of the first embodiment of the present application. Figure 1 The first embodiment of the present application provides a big data-based early monitoring system for pancreatic cancer and is described in detail.

[0038] The pancreatic cancer early monitoring system based on big data includes a data acquisition module 101 , a feature extraction module 102 , a risk assessment module 103 and an early warning trigger module 104 .

[0039] The data acquisition module 101 is used to obtain the user's laboratory test data, imaging test data and clinical symptom data, wherein the laboratory test data includes serum CA19-9 level, serum CEA level and blood glucose level, the imaging test data includes pancreatic parenchyma density value and pancreatic duct dilatation degree data, and the clinical symptom data includes abdominal pain degree, weight change and jaundice index.

[0040] The data acquisition module 101 is a basic component of the system. Its function is to comprehensively and accurately collect the user's health data, including laboratory test data, imaging test data and clinical symptom data, to support subsequent feature extraction, risk assessment and early warning operations. The data acquisition module 101 needs to have a variety of data acquisition interfaces, real-time monitoring capabilities and high-precision data processing capabilities to ensure the integrity and accuracy of the collected data.

[0041] In terms of the collection of laboratory test data, module 101 obtains key indicators such as serum CA19-9 level, serum CEA level and blood glucose level by connecting to laboratory testing equipment or database. These indicators are usually obtained through quantitative analysis instruments, such as enzyme-linked immunosorbent assay (ELISA) to measure CA19-9 and CEA levels, and glucose oxidase method to measure blood glucose level. Module 101 needs to be able to receive this data in real time, or update the data through a periodic interface. In addition, the module should have the function of identifying abnormal data, for example, when an indicator is higher or lower than a predetermined reference range, a mark is generated to prompt the subsequent processing module.

[0042] The acquisition of imaging examination data is completed by module 101 through docking with imaging equipment (such as CT scanner, MRI or ultrasound equipment). Specifically, the module obtains the density value of pancreatic parenchyma and the degree of pancreatic duct dilatation data. These data can be extracted from the image data through image segmentation algorithms, such as marking the pancreatic area through automated image analysis software, measuring its density value and quantifying the diameter and morphological characteristics of pancreatic duct dilatation. In order to improve the timeliness and practicality of the data, module 101 can integrate the image database to directly retrieve the patient's historical image data and support time series analysis.

[0043] Clinical symptom data is mainly collected through patient self-report, questionnaires or wearable device records. Module 101 captures the user's reported abdominal pain level, weight change and jaundice index, and realizes data input through interaction with the user terminal (such as a mobile device or a health management platform). The degree of abdominal pain can be quantified by the patient's selection of a preset pain level scale (such as 0 to 10 points), weight change data can be obtained through an interface with a smart scale, and the jaundice index is evaluated by combining laboratory data (such as bilirubin levels) or image analysis (such as skin or eye color). In addition, the module must also support clinical physician input, allowing physicians to supplement or correct the user's symptom data.

[0044] The data collection module 101 needs to have a powerful data preprocessing function to ensure the quality and consistency of the collected data. This includes data format standardization, missing value filling and outlier filtering. For example, for the collected multimodal data, the module 101 can convert it into a unified data format (such as JSON or CSV) for storage, and use an interpolation algorithm to process the discontinuous data that may be generated during the collection process.

[0045] Module 101 also supports extended functions, including dynamic data collection and remote monitoring. The dynamic data collection function allows the module to adjust the collection frequency according to changes in the user's health status. For example, when laboratory data or symptom data indicate that there may be a high risk, the module can automatically increase the collection frequency of imaging examination data to capture the development of the disease in a more timely manner. The remote monitoring function is implemented through the Internet of Things (IoT) technology, allowing the module to collect data in real time from distributed devices, such as obtaining relevant information from the user's home testing equipment or wearable smart devices.

[0046] Through the integration and optimization of these functions, the data acquisition module 101 not only ensures the breadth and depth of the data, but also provides reliable basic support for subsequent modules, comprehensively improving the performance and practicality of the early pancreatic cancer monitoring system.

[0047] Furthermore, the data acquisition module is specifically used for:

[0048] Time series features were generated based on the change rate of serum CA19-9 levels in laboratory examination data collected at consecutive time points and the dynamic changes in the degree of pancreatic duct dilatation in imaging examination data.

[0049] In this embodiment, the data collection module not only realizes the collection of routine health data, but also develops a time series feature generation method specifically for the dynamic changes of laboratory test data and imaging test data. This method aims to provide more diagnostic information by capturing the changing patterns of data at different time points, thereby significantly improving the sensitivity and accuracy of early monitoring of pancreatic cancer.

[0050] First, the data acquisition module dynamically monitors the changes in serum CA19-9 levels at continuous time points. This process requires the collection of CA19-9 values ​​from multiple laboratory tests by the user, which may show complex change patterns due to the patient's health status, lifestyle or treatment intervention. The module records this data at predetermined time intervals, such as once a week or a month, to ensure that enough time points are captured to reflect its dynamic changes. After the collection is completed, the module calculates the difference between each two adjacent time points and normalizes it with a specific time span (such as day, week or month) to generate rate of change data. This rate of change not only reflects the direction and magnitude of the increase or decrease in the value, but also provides trend information reflecting the possibility of pancreatic lesions, such as continued increases or large fluctuations. In addition, to improve the robustness of the analysis, the module also has an outlier detection function, such as using statistical methods to exclude abnormal readings caused by experimental errors or one-time external interference.

[0051] Secondly, the data acquisition module captures the dynamic changes in the degree of pancreatic duct dilatation by docking with imaging equipment. Unlike the numerical changes in laboratory data, the processing of imaging data is more complicated because it requires the extraction of structural features from the original images. The module is integrated with the image analysis algorithm to first preprocess the pancreatic CT or MRI images of the pancreas at consecutive time points, including denoising, segmentation, and feature extraction. When extracting the features of pancreatic duct dilatation, the module not only records the diameter value of the pancreatic duct, but also analyzes the morphological changes of the pancreatic duct, such as whether the dilatation range extends along the main axis of the pancreatic duct and whether there are significant irregularities or multi-node dilatation patterns. For the feature data at these time points, the module uses time series analysis techniques, such as statistical calculations based on sliding windows, to generate time series features that describe dynamic changes, including the expansion speed, direction of change, and frequency of morphological abnormalities.

[0052] In the process of combining laboratory data with imaging data, the data acquisition module pays special attention to the time alignment between different modal data. For example, the detection frequency of serum CA19-9 levels may be much higher than the frequency of imaging examinations, so the module designs a multi-step interpolation mechanism to establish a time correspondence between the two. For cases where time points do not completely overlap, the module generates predicted values ​​based on the data at the nearest time point or through trend extrapolation to ensure the integrity and consistency of dynamic features.

[0053] The generated time series features are not limited to simple values ​​or rates of change, but also include comprehensive features, such as the association between the volatility of laboratory data and the dynamic expansion pattern of imaging data. The data acquisition module uses the fusion analysis function to cross-correlate the dynamic features of the two modalities and find key patterns that may indicate the risk of pancreatic cancer. For example, when the rate of change of CA19-9 levels shows a significant positive correlation with the rate of pancreatic duct expansion, the module generates a risk marker and attaches the marker to the time series features for further analysis by subsequent modules.

[0054] In order to support a wider range of application scenarios, the data acquisition module also provides user customization functions, allowing the frequency of data acquisition and algorithm parameters for dynamic feature generation to be adjusted according to specific diagnostic needs. For example, in high-risk patients, the module can increase the frequency of data acquisition or refine the granularity of change rate calculation to obtain a more accurate dynamic change pattern. For low-risk patients, the module can reduce the complexity of data processing by reducing the acquisition frequency or simplifying the analysis model, while still maintaining the integrity of key features.

[0055] The feature extraction module 102 is connected to the data acquisition module and is used to extract feature parameters from the acquired data and construct a feature vector, wherein the feature vector includes a biochemical index feature sub-vector, an image feature sub-vector and a symptom feature sub-vector.

[0056] The feature extraction module 102 is one of the core components of the system. Its function is to extract key features from the multimodal data provided by the data acquisition module 101 and construct a feature vector that can be used for risk assessment. The feature extraction module needs to have a high degree of flexibility and adaptability to process three completely different data types, namely laboratory test data, imaging test data, and clinical symptom data, and ultimately generate a biochemical index feature subvector, an imaging feature subvector, and a symptom feature subvector. The implementation process of the module is described in detail below.

[0057] For the processing of laboratory test data, the feature extraction module first receives key indicators such as serum CA19-9 levels, serum CEA levels, and blood glucose levels. These indicators may have different collection time points and unit formats, so the module first performs data cleaning and standardization, such as unifying units and timelines. After that, the module extracts features for each indicator, including basic statistical features (such as mean, maximum, minimum, standard deviation) and dynamic features (such as rate of change, trend of change). In order to enhance the diagnostic value of the features, the module can also combine historical data to identify specific patterns that may indicate the risk of pancreatic cancer through anomaly detection algorithms, such as a sustained increase in CA19-9 levels or significant fluctuations in blood glucose levels.

[0058] In the processing of imaging examination data, the module receives features such as pancreatic parenchymal density values ​​and the degree of pancreatic duct dilatation. The module first calls the image analysis algorithm to process the original image data, including noise removal, region segmentation, and target detection. Specifically, the module accurately locates the pancreatic region through the segmentation algorithm and extracts the distribution characteristics of the pancreatic parenchymal density value, such as the mean, variance, and skewness of the density. At the same time, for the degree of pancreatic duct dilatation, the module uses edge detection and morphological analysis methods to measure the length, width, and irregularity of the pancreatic duct and its dilatation. In addition, the module can align and compare image data at multiple time points to capture the dynamic changes in pancreatic structure or density over time to provide additional diagnostic information.

[0059] For the processing of clinical symptom data, the module extracts features from the degree of abdominal pain, weight change, and jaundice index reported by users or collected by wearable devices. The module first classifies the symptom data, for example, the degree of abdominal pain is divided into mild, moderate, and severe according to the pain score. Next, the module calculates the dynamic changes in the frequency, duration, and severity of the symptoms. For example, the module can detect nonlinear trends in weight changes and identify significant weight loss patterns associated with pancreatic cancer. In addition, for the jaundice index, the module can generate composite features by combining laboratory test results (such as bilirubin levels) with the user's self-reported symptoms to further improve the diagnostic value of the features.

[0060] The feature extraction module adopts a modular architecture in the above three feature extraction processes so that the feature extraction method can be flexibly adjusted according to the needs. For example, for the examination data of different patients, the module can dynamically select the most suitable feature extraction algorithm or strategy. In addition, the module supports feature dimensionality reduction operations, such as reducing feature dimensions through principal component analysis (PCA) or autoencoder models, to ensure that the generated feature vector contains rich diagnostic information without causing computational burden on the subsequent risk assessment module.

[0061] By implementing the above steps, the feature extraction module 102 can comprehensively mine the key features in the laboratory test, imaging examination and clinical symptom data, and convert them into feature vectors that can be used in deep learning models, providing strong support for subsequent risk assessment modules and early warning trigger modules.

[0062] Furthermore, the feature extraction module is specifically used for:

[0063] Receive continuous time point data of serum CA19-9 level, serum CEA level and blood glucose level, and normalize the data of each indicator to eliminate the influence of different dimensions;

[0064] Based on the normalized data of multiple time points, calculate the change trend and fluctuation range of each indicator, and identify abnormal feature points that exceed the preset threshold;

[0065] The distribution pattern of the abnormal feature points is compared with historical data to generate a set of time series feature values ​​as the basis of the biochemical indicator feature sub-vector.

[0066] First, the feature extraction module receives the user's serum CA19-9 level, serum CEA level, and blood glucose level data at different time points. These data may come from multiple laboratory test results, and usually contain characteristics such as a wide range of values, inconsistent units, and different collection time intervals. The first step of the module is to normalize all received data to eliminate the impact caused by dimensional differences between different indicators. This process converts the original data into a standard normal distribution form by calculating the mean and standard deviation of each indicator, thereby ensuring that all indicators have the same numerical range and statistical characteristics in subsequent analysis.

[0067] After normalization, the module calculates the trend and fluctuation range of each indicator based on the normalized data at multiple time points. The trend calculation identifies the growth or decline of the indicator by analyzing the time series behavior of the data. For example, the module can apply a linear fitting method to calculate the trend slope to quantify the rate of change of the indicator over a period of time. The fluctuation range analysis focuses on the fluctuation amplitude of the indicator in the time series, and evaluates whether these fluctuations are periodic or abnormal by calculating the difference between the maximum and minimum values ​​and combining time weighting.

[0068] Next, the module further processes the data generated in the above analysis to identify abnormal feature points that exceed the preset threshold. The identification of abnormal feature points not only depends on the absolute value of the data at a single time point, but also requires a comprehensive judgment based on historical trends and fluctuation characteristics. For example, if the serum CA19-9 level at a certain time point is significantly higher than the set reference range and deviates significantly from the previous trend of change, the module will mark the point as an abnormal feature point. In addition, the module can distinguish specific abnormal patterns that may be related to pancreatic cancer, such as rapidly increasing CA19-9 levels or long-term hyperglycemia, by automatically classifying the data.

[0069] After identifying abnormal feature points, the module will compare the distribution patterns of these abnormal points with historical data to discover deeper diagnostic information. This comparison process involves pattern matching analysis of time series. The module will use historical data to build a multidimensional feature model and project the time distribution of the current abnormal point into this model to evaluate whether the current feature is consistent with the known pancreatic cancer feature pattern. For example, the module may find that the distribution of abnormal points in the current data is similar to the known high-risk feature pattern of pancreatic cancer, thereby improving the reliability of diagnosis.

[0070] Finally, the module generates a set of time series eigenvalues ​​based on the above analysis. These eigenvalues ​​integrate the dynamic change information of trends, fluctuations and outliers, forming the basis of the characteristic sub-vectors of biochemical indicators. These time series eigenvalues ​​can be designed as vectors of fixed length to facilitate the input processing of subsequent deep learning models. At the same time, the module supports dynamic expansion of different patient data. Even if data is missing at a specific time point, complete time series features can still be generated through interpolation and prediction algorithms.

[0071] Furthermore, the feature extraction module is specifically used for:

[0072] Receiving data on the density value of the pancreatic parenchyma and the degree of dilatation of the pancreatic duct, and preprocessing the image data, including denoising and segmenting the target area;

[0073] Extract pancreatic edge structure and pancreatic duct morphological features from the segmented image region, including the length, diameter and curvature of the pancreatic duct;

[0074] The extracted morphological features are combined with the pancreatic density value to construct the image feature sub-vector.

[0075] After receiving the original image data of pancreatic parenchymal density and pancreatic duct dilatation, the feature extraction module first preprocesses these image data. The first task of preprocessing is denoising. This step uses a variety of image processing algorithms to remove background noise that may interfere with feature extraction, such as reducing random noise in the image through Gaussian filters, or separating high-frequency noise components in the image through wavelet transforms. In addition, for possible motion artifacts or scanning artifacts in pancreatic images, the module also uses frequency domain filtering-based artifact removal technology to ensure the clarity and data integrity of the target area.

[0076] Image segmentation is the core part of the preprocessing process. The module automatically segments the pancreatic region in the image through advanced segmentation algorithms. For example, it uses a semantic segmentation model based on deep learning to accurately extract the pancreatic parenchyma and pancreatic duct area from a complex background. This segmentation algorithm can be achieved by training a large amount of labeled image data. Its core is to extract spatial features from the image in combination with a convolutional neural network (CNN). At the same time, it improves the accuracy of segmentation by introducing morphological prior knowledge, such as the typical morphology and position of the pancreas. After segmentation, the module can also automatically evaluate the quality of the segmented area to ensure that the segmentation results can support the subsequent feature extraction steps.

[0077] From the segmented pancreatic region, the module further extracts the morphological features of the pancreatic edge structure and the pancreatic duct. When extracting the length and diameter of the pancreatic duct, the module accurately marks the main axis of the pancreatic duct through the centerline extraction algorithm, and calculates the local diameter value at each node of the main axis to obtain the distribution of the length and diameter of the pancreatic duct. In addition, the module evaluates the curvature changes of the pancreatic duct through the curvature analysis method to capture abnormal morphologies that may indicate lesions, such as local severe bending of the pancreatic duct or irregular expansion of branches. For the extraction of pancreatic edge structures, the module uses a contour detection algorithm based on gradient analysis, which can identify the edge morphology of the pancreas, analyze its smoothness and continuity, and detect possible abnormal protrusions or irregular edge features.

[0078] The innovation of the feature extraction module is that it not only analyzes the above-mentioned extracted morphological features independently, but also combines them with the pancreatic parenchyma density values ​​to construct a comprehensive image feature sub-vector. The fusion of density values ​​is completed by dividing the pancreatic region into multiple sub-regions and calculating the average density, density distribution range and density gradient of each sub-region. The module designs a fusion method based on spatial weights to assign weights to the density features of different sub-regions to highlight the diagnostic value of key areas of the pancreas. For example, areas with significant changes in density gradients may indicate early signs of tumor formation.

[0079] Finally, the construction of the image feature sub-vector integrates morphological features and density features to form a multi-dimensional vector of fixed length. These vectors not only cover the comprehensive information of pancreatic edge structure, pancreatic duct morphological features and density distribution, but can also be further optimized through the module's built-in dimensionality reduction algorithm to ensure that the data input into the subsequent risk assessment module is both concise and efficient. In addition, the module supports personalized analysis of imaging data of different patients, such as increasing the granularity of pancreatic duct segmentation feature extraction for high-risk patients to capture more subtle morphological changes.

[0080] Furthermore, the feature extraction module is specifically used for:

[0081] Receive the user's self-reported data on abdominal pain, weight change, and jaundice index, and classify the data to distinguish between acute and chronic symptoms;

[0082] For acute symptoms, assess their frequency and duration; for chronic symptoms, analyze their cumulative change trends;

[0083] The symptom severity score is compared with the user's previous health records to generate a symptom feature subvector.

[0084] After receiving the user's self-reported data on abdominal pain, weight change, and jaundice index, the feature extraction module first classifies the data into acute and chronic symptoms. This classification process is defined by the time of occurrence and duration of the symptoms. For example, the score of abdominal pain is usually expressed by a quantitative numerical level selected by the user (such as 0 to 10), and acute abdominal pain is defined as sudden and intense symptoms in a short period of time; in contrast, chronic abdominal pain manifests as long-term mild to moderate pain, and the symptoms may occur periodically. The module uses a built-in rule library to automatically classify the data reported by the user, while allowing the user to adjust the classification results according to their actual feelings to ensure the accuracy and personalization of the classification.

[0085] After the classification is completed, the module further analyzes the acute symptoms, focusing on evaluating their frequency and duration. For example, for the analysis of acute abdominal pain, the module will count the number of symptoms reported by the user within a specific time period (such as a week or a month) and calculate the duration of each abdominal pain. In this way, the module can capture the temporal distribution characteristics of acute symptoms, such as whether there is a high incidence period of symptoms or a trend of gradually increasing frequency. For the processing of the acute jaundice index, the module will combine the user's skin and eye color change records to evaluate the time point and duration of jaundice onset, thereby providing a reference for subsequent risk assessment.

[0086] When dealing with chronic symptoms, the module focuses more on their cumulative change trends. For example, for weight change data, the module will record the user's weight data at multiple time points, and analyze whether there is a significant gradual decline or fluctuation feature by calculating its long-term trend line. For chronic jaundice, the module will calculate the user's average jaundice index, and combine its fluctuation range to determine whether there is a long-term high trend. Through this analysis of cumulative change trends, the module can provide a more comprehensive description of symptom characteristics from a time series perspective, capturing the potential threat of chronic symptoms to health.

[0087] After the feature analysis of acute and chronic symptoms is completed, the module compares the severity score of the symptoms with the user's previous health records to generate more targeted diagnostic features. The module extracts reference indicators related to the current symptoms from the user's historical health data, such as previous weight change trends or the fluctuation range of the early jaundice index. By comparing current symptom data with historical records, the module is able to identify abnormal patterns. For example, if the user's currently reported abdominal pain score is significantly higher than its long-term average and the frequency of occurrence increases, the module will mark this symptom as a high-risk feature. In addition, for some abnormal symptoms that are inconsistent with previous health records, the module will generate additional labels for further processing by subsequent modules.

[0088] Finally, the module generates a symptom feature subvector based on the above analysis. This subvector is a high-dimensional feature representation that integrates symptom classification, temporal changes, and historical comparisons, and can accurately reflect the potential correlation between the user's current symptoms and pancreatic cancer risk. The subvector is constructed using the module's built-in feature embedding technology to map multiple symptom features into a fixed-dimensional vector space, ensuring consistency in data structure with other feature subvectors (such as biochemical indicators and imaging feature subvectors), thereby facilitating the processing and integration of subsequent deep learning models.

[0089] The feature extraction module significantly enhances the diagnostic value of symptom data, not only capturing short-term high-risk signals of acute symptoms, but also identifying the long-term cumulative effects of chronic symptoms, and combining this information with the user's health history to provide more comprehensive and personalized risk assessment support.

[0090] The risk assessment module 103 is connected to the feature extraction module and is used to obtain the pancreatic cancer risk probability based on the feature vector using a pre-trained deep learning model.

[0091] The risk assessment module 103 is one of the core components of the system. Its function is to receive the feature vectors generated by the feature extraction module 102, analyze and process these features through the pre-trained deep learning model, and thus calculate the probability value of pancreatic cancer risk. The design of this module not only needs to ensure efficient processing and accurate assessment of input features, but also needs to have the ability to dynamically adapt to the characteristics of different patient data to achieve flexible and reliable risk prediction.

[0092] The risk assessment module first receives feature vectors containing multimodal data, including biochemical indicator feature subvectors, imaging feature subvectors, and symptom feature subvectors. These feature vectors undergo necessary preprocessing before entering the model, such as dimensionality checking, numerical normalization, and missing value completion, to ensure that the input data meets the requirements of the deep learning model. Normalization eliminates the interference of different data types and dimensions on the model evaluation results by mapping the feature values ​​to a fixed range, such as between 0 and 1.

[0093] A pre-trained deep learning model is used inside the module to calculate the risk probability. The model can be a multi-layer perceptron (MLP), convolutional neural network (CNN), recurrent neural network (RNN) or multimodal fusion network. The model structure is carefully designed to fully utilize the complementarity of multimodal features. For example, biochemical indicator features usually have strong temporal dependencies and can be modeled by recurrent neural networks; imaging features have spatial correlations and deep features can be extracted by convolutional neural networks; and symptom features are more suitable for processing by fully connected networks to capture complex nonlinear relationships. Finally, these features are integrated through a fusion layer to form a high-dimensional representation of a comprehensive assessment.

[0094] During the inference phase of the model, the risk assessment module inputs the feature vector into the deep learning model, and after a series of nonlinear transformations and weight adjustments, it gradually calculates the output value representing the risk of pancreatic cancer. This output value is usually a probability between 0 and 1, indicating the possibility of the user suffering from pancreatic cancer. To ensure the accuracy and robustness of the assessment, the module also applies multiple optimization strategies, such as overfitting control based on weight regularization, training optimization based on dynamic learning rate, and multi-model fusion strategy during model inference.

[0095] In addition, in order to improve the personalization and dynamic adaptability of risk assessment, the module can use the patient's historical data to dynamically adjust the model's reasoning results. For example, when it is detected that certain feature values ​​continue to fluctuate within a specific range, the module will assign higher weights to these features to strengthen the model's sensitivity to abnormal signals. At the same time, the risk assessment module supports real-time fine-tuning of the model through an online learning mechanism to adapt to newly added data features, thereby improving the assessment capabilities for different patients and special cases.

[0096] The risk assessment module also has log recording and traceability functions. The input, output and key intermediate steps of each model reasoning will be recorded to ensure that the entire assessment process is transparent and traceable. These log data can serve as an important basis for subsequent model optimization and diagnostic verification.

[0097] Through the above functions, the risk assessment module 103 can not only generate accurate pancreatic cancer risk probability for the user, but also provide highly personalized and dynamically adaptive assessment services based on multimodal data.

[0098] Furthermore, the deep learning model includes a feature fusion network layer, a temporal feature extraction layer, and a risk prediction layer;

[0099] The feature fusion network layer is used to receive and process input features of different modalities, including a first feature processing unit, which is used to reduce the dimension and extract features of the biochemical index feature subvectors through three fully connected layers, and use a nonlinear activation function to optimize the expression of features to obtain a biochemical feature representation; a second feature processing unit, which is used to extract the spatial feature relationship of the image data through a convolution operation, and generate an image feature representation through a fully connected layer; a third feature processing unit, which is used to perform feature mapping on the symptom feature subvector to obtain a symptom feature representation; a dynamic weighted fusion unit, which is used to receive the biochemical feature representation, the image feature representation, and the symptom feature representation; and based on the received data, calculate the feature weights using an attention mechanism to generate a fused feature vector;

[0100] The temporal feature extraction layer is used to analyze the temporal evolution law of the fused feature vector, including a temporal data cache unit, which is used to maintain a set of historical fused feature vectors and construct a temporal feature matrix; a gated recurrent unit, which is used to analyze the temporal evolution law of the temporal feature matrix and generate a hidden state vector; and a multi-head attention mechanism, which is used to enhance the features of the hidden state vector and generate an enhanced temporal feature vector.

[0101] The risk prediction layer is used to generate the final risk assessment result, including a feature aggregation unit, which is used to receive the fused feature vector and the enhanced time series feature vector, realize feature merging through a skip connection structure, and generate a comprehensive feature vector using a dimensionality reduction operation; a risk probability calculation unit, which is used to perform feature mapping on the comprehensive feature vector through multiple fully connected layers, and output the pancreatic cancer risk probability distribution using a Softmax function.

[0102] The deep learning model is the core technical component for realizing the early monitoring system of pancreatic cancer. The design includes feature fusion network layer, time series feature extraction layer and risk prediction layer. It aims to achieve efficient assessment and early warning of pancreatic cancer risk through multimodal data fusion, time series analysis and accurate prediction.

[0103] The feature fusion network layer is used to receive and process multimodal input features from the data acquisition module and the feature extraction module, including biochemical index feature subvectors, image feature subvectors, and symptom feature subvectors. In view of the features of different modalities, this layer is designed with three feature processing units, each of which uses a processing method that is suitable for the characteristics of the input features. The first feature processing unit reduces the dimension and extracts features of the biochemical index feature subvectors through three fully connected layers. Each fully connected layer uses a nonlinear activation function, such as ReLU or Leaky ReLU, to improve the feature expression ability and nonlinear modeling ability, and finally generates a biochemical feature representation. The second feature processing unit uses convolution operations to extract the spatial correlation features of the data based on the spatial characteristics of the image data, such as the local correlation of the pancreatic density value and the pancreatic duct morphology, and then generates an image feature representation through a fully connected layer. The third feature processing unit performs feature mapping on the symptom feature subvector, maps the discretized symptom data (such as pain level and weight change) into a continuous feature space, and generates a symptom feature representation. To further fuse the features of these modalities, the dynamic weighted fusion unit introduces an attention mechanism, dynamically adjusts the weight of each feature according to the importance of each feature in different risk assessment scenarios, and generates a unified fused feature vector through weighted summation.

[0104] The time series feature extraction layer is responsible for analyzing the evolution of the fused feature vector in the time dimension to capture the pattern and potential trend of the feature changes over time. This layer first maintains a set of historical fused feature vectors through the time series data cache unit, and constructs a time series feature matrix containing features of multiple time points to ensure the time continuity and integrity of the data. Subsequently, the gated recurrent unit (GRU) analyzes the feature changes at these time points, effectively processes long-term and short-term dependencies through its gating mechanism, and generates a hidden state vector that can reflect the deep pattern of the feature evolution over time. In order to further enhance the expressive power of time series features, a multi-head attention mechanism is applied to the hidden state vector. Multiple attention heads are used to capture the correlation between different time steps, generate enhanced time series feature vectors, and provide higher feature modeling capabilities.

[0105] After receiving the fused feature vector and the enhanced time series feature vector, the risk prediction layer effectively merges the two through the feature aggregation unit. The feature aggregation unit uses the skip connection structure to integrate the advantages of the fused features and the time series features into a comprehensive feature vector while maintaining the original feature information. Subsequently, the comprehensive feature vector reduces redundant features through dimensionality reduction operations while retaining key information, providing optimized input for subsequent risk assessment. Based on the comprehensive feature vector, the risk probability calculation unit uses multiple fully connected layers to gradually extract high-order features, and converts the output value into a probability distribution through the Softmax function, and finally generates the risk probability of pancreatic cancer. This probability distribution usually includes two possibilities: healthy and sick, and its value reflects the relative risk level of the user suffering from pancreatic cancer.

[0106] This deep learning model can capture complex feature relationships and dynamic change patterns through multi-level, multi-modal feature extraction, time series analysis and risk prediction, providing highly accurate pancreatic cancer risk assessment capabilities.

[0107] Furthermore, the dynamic weighted fusion unit is specifically used for:

[0108] Receiving biochemical signatures b , image feature representation F i , Symptoms and characteristics s , and map them to the unified feature space through nonlinear functions through the following formula (1):

[0109] T k =tanh(W k ·F k +b k ),k∈{b,i,s} (1)

[0110] Among them, W k is the feature projection matrix; b k is the bias vector; T k Indicates the biochemical signature received F b , image feature representation F i , Symptoms and characteristics s The unified feature representation of any feature representation in , where T b is a unified feature representation of biochemical features, T i is the unified feature representation of image features, T s It is a unified feature representation of the symptom characteristics;

[0111] According to the following formula (2), the fusion feature vector F is calculated c :

[0112]

[0113] Where λ is the adjustment factor; where angle φ k Calculate according to the following formula (3):

[0114]

[0115] Among them, T m is the feature vector, T m =T b +T i +T s .

[0116] The dynamic weighted fusion unit of this embodiment aims to generate a fusion feature vector F that can comprehensively reflect multi-modal feature information by uniformly mapping and weighted fusion of different modal features. c .

[0117] Formula (1) is used to express the biochemical characteristics F b , image feature representation F i and symptom characteristics indicate F s Mapped to a unified feature space. The mapping process is completed by the nonlinear function tanh, where W k is the feature projection matrix, which defines the linear transformation rule of each feature from the input space to the unified feature space; b k is a bias vector used to adjust the benchmark of the mapping result. Specifically, W k The dimension of F is determined by the dimension of the input features and the dimension of the unified feature space. b is 64-dimensional, and the unified feature space is 128-dimensional, then W b The size of is 128 × 64. The bias vector b k The size of is consistent with the dimensionality of the unified feature space, which is 128 in the above example.

[0118] The role of the nonlinear function tanh is to compress the eigenvalue range to between -1 and 1, while enhancing the nonlinear expression ability, so as to better reflect the complex feature relationship. The result after mapping T k represents the feature vector in the unified feature space, where T b T is a unified representation of biochemical characteristics. i is a unified representation of image features, T s A unified representation of symptom characteristics.

[0119] Formula (2) is used to generate the fusion feature vector F c, which combines the contributions of multimodal features. The λ in the formula is an adjustment factor used to control the overall amplitude or importance of the fused features. Its value is usually tuned according to the distribution of experimental data. The reference value is λ=0.5 to λ=1.5. The specific value can be determined by the feature distribution of the validation dataset.

[0120] The weight is calculated from the angle φ k and the normalization factor, where the angle φ k Represents the feature T k and the modal overall characteristic T m The angle between them defines T k The relative directionality in the overall feature. The calculation formula (3) calculates the similarity of two vectors through the inner product and normalizes it with the modulus length to get the cosine value of the angle. It is further converted into an angle value through the inverse cosine function. Here, T m =T b +T i +T s It is the sum of all modal characteristics and represents the overall direction and characteristics of the mode. k The range is from 0 to π, indicating a directional relationship from completely consistent to completely opposite.

[0121] Formula (2) is improved by introducing cos(φ k ) as weights, which strengthens the importance of modal features that are closer to the overall feature direction. Weight normalization ensures that the sum of the contributions of multimodal features is 1, thereby avoiding the dominant effect of a single modal feature. This directionality-based weighting method enhances the physical meaning and robustness of feature fusion, and is particularly suitable for scenarios where modal feature expressions have significant directionality.

[0122] Fusion feature vector F c The calculation logic of can take into account the independent contribution of each modal feature and its synergy with the overall feature. For example, when a modal feature is m occupies a large proportion (i.e. cos(φ k ) is large), its corresponding weighted contribution also increases accordingly, so that the fused feature vector can more comprehensively reflect the importance of this modal feature.

[0123] Furthermore, the multi-head attention mechanism in the temporal feature extraction layer realizes feature enhancement of the hidden state vector through the following steps:

[0124] The hidden state vector output by the gated loop unit Decompose into multiple sub-vector groups according to the time step T and feature dimension d, corresponding to each attention head k, where d k =d / m represents the feature dimension of attention head k, and m is the number of attention heads;

[0125] For each attention head k, the query matrix Q is generated according to the following formula (4): k , key matrix K k Sum value matrix V k :

[0126]

[0127] in, is the weight matrix,

[0128] The attention weight matrix A is calculated according to the following formula (5): k :

[0129]

[0130] in, represents the average value of the key matrix, which is used to reflect the characteristic deviation between time steps; γ is the adjustment coefficient;

[0131] According to the following formula (6), the output O of attention head k is generated k :

[0132] Using attention weight A k Pair matrix V k Perform weighted summation to generate the output O of each head k :

[0133] O k =A k ·V k (6)

[0134] Among them, V k is the value matrix;

[0135] According to the following formula (7), the outputs of all attention heads (O1, O2, ..., O m ) is combined into an enhanced time series feature vector F through concatenation operation e :

[0136] F e =Concat(O1,O2,…,O m )·W o (7)

[0137] in, is the linear transformation matrix used to map to the final output dimension.

[0138] The time series feature extraction layer in this embodiment enhances the feature expression capability of the hidden state vector through a multi-head attention mechanism, thereby more accurately capturing the complex relationship between time series features.

[0139] The hidden state vector at the output of the GRU In the above equation, T represents the number of time steps and d represents the feature dimension. First, H is decomposed into multiple sub-vector groups according to the time step and feature dimension, corresponding to each attention head. The number of attention heads is m, and the feature dimension of each head is d. k According to the formula d k =d / m. This decomposition method can parallelize multiple attention calculations, thereby improving the efficiency and depth of temporal feature extraction.

[0140] For each attention head k, the query matrix Q is generated by formula (4) k , key matrix K k Sum value matrix V k The calculation of these matrices depends on the weight matrix and The dimension of each weight matrix is These weight matrices are trainable parameters that are optimized during the training process of the deep learning model and are used to achieve a linear mapping from the input hidden state vector to the representation of a specific attention head. For example, if the input hidden state vector dimension is d=128, the feature dimension of each head is d k =16, then and The size of the matrix Q is 128 × 16. k , K k 、V k Represent the query vector, key vector, and value vector of the hidden state in the current time step respectively.

[0141] Formula (5) is used to calculate the attention weight matrix A k The core is to measure the query matrix Q through dot product operation and normalization function. k With the key matrix K k The formula introduces the feature deviation between time steps in is the average of the key matrix, representing the global center of features at all time steps. The deviation term Δ k The purpose of is to enhance the sensitivity to feature changes between time steps, and to control its contribution to attention calculation by adjusting the coefficient γ. The recommended initial value of γ is 0.1 to 1, and the specific value can be tuned through the training and validation data set. The normalization of the weight matrix is ​​completed through the Softmax function to ensure that the sum of all time step weights is 1, thereby ensuring the physical meaning of attention distribution.

[0142] The denominators in formula (5) are used to calculate the attention weight matrix A kIt is the core part of the process, and its function is to normalize the features between different time steps to ensure that the weight values ​​are statistically significant and can reflect the importance of each time step.

[0143] The main content in the denominator is a cumulative term The purpose of this accumulation term is to normalize the feature correlation of all time steps j, ensuring that the attention weight A of each time step k Falls within a certain range (usually 0 to 1), and the sum of all weights is 1. The specific description is as follows:

[0144]

[0145] This part is the query matrix Q k The j-th time step with the key matrix The dot product of k The similarity between the features of the jth time step. The larger the dot product result, the higher the Q k and The closer the directions in the feature space are, the greater the contribution of the current time step to the query feature.

[0146] γ·diag(Δ j )

[0147] Adjustment term γ·diag(Δ j ) is used to introduce the time-step deviation information of the feature. Here in is the average value of the key matrix K, which represents the global center value of the key features of all time steps. Deviation Δ j The offset of the key feature at the jth time step relative to the overall center is represented by diag(Δ j ) generates a diagonal matrix and weights it element-wise.

[0148] Formula (6) uses the attention weight matrix A k Pair matrix V k Perform weighted summation to generate the output O of attention head k k The result of the weighted summation combines the value vector of each time step with its corresponding attention weight to generate a feature representation that contains temporal correlation. Through this step, each attention head captures the correlation and information interaction between different time steps.

[0149] Formula (7) is the output of all attention heads O1, O2, ..., O m Through concatenation operation, it is combined into an enhanced time series feature vector F e The concatenation operation stacks the output of each head in the column direction to form a higher-dimensional feature representation with a size of Then, through the linear transformation matrix Map the concatenated features back to the final output dimension d. Linear transformation matrix W o The function of is to integrate the outputs of multiple attention heads and reduce the dimensionality of high-dimensional features to adapt to the input requirements of subsequent modules.

[0150] Through this design, the multi-head attention mechanism can effectively capture the complex dependencies and multi-level feature interactions in time series features. Compared with the existing technology, this method not only improves the accuracy of time series feature extraction, but also significantly enhances the model's adaptability to dynamic feature changes through feature deviation and multi-head parallel mechanism.

[0151] Furthermore, the risk probability calculation unit in the risk prediction layer is specifically used to:

[0152] Receive the comprehensive feature vector F output by the feature aggregation unit c ,in, d c represents the dimension of the comprehensive feature vector;

[0153] According to the following formula (8), the comprehensive feature vector F is activated by the feature activation function g(x) c Perform nonlinear mapping to generate the feature vector F after preliminary processing a :

[0154] F a =g(W1·F c +b1) (8)

[0155] Among them, W1 is the feature mapping matrix, d h is the feature dimension of the hidden layer; b1 is the bias vector, g(x) is the activation function, which is implemented using the following formula (9):

[0156]

[0157] According to the following formula (10), the risk feature vector F is generated r :

[0158] F r =W2·(F a ⊙tanh(W t ·F a )+b2 (10)

[0159] Among them, W2 is the risk characteristic weight matrix, d r is the dimension of the risk feature vector; W tGenerate the matrix for adaptive weights, ⊙ represents element-by-element multiplication; b2 is the bias vector,

[0160] According to the following formula (11), the pancreatic cancer risk probability distribution P is generated:

[0161] P = Softmax(W3·F r +b3) (11)

[0162] Among them, W3 is the weight matrix, b3 is the bias vector, Softmax is the activation function.

[0163] The risk probability calculation unit of this embodiment is located in the risk prediction layer and is a key component for realizing pancreatic cancer risk assessment. Its design is based on nonlinear mapping, multi-layer feature transformation and probability distribution generation of deep learning, and can convert the input comprehensive feature vector F into the final pancreatic cancer risk probability distribution P.

[0164] Comprehensive feature vector F c is the input vector generated by the feature aggregation unit, with dimension d c This vector contains the integrated information from the fused feature vector and the enhanced time series feature vector, and is the final representation of the multimodal data. The risk probability calculation unit first uses formula (8) to calculate F c Perform nonlinear mapping to generate the feature vector F after preliminary processing a .

[0165] The dimension of the feature mapping matrix W1 in formula (8) is Used to convert the comprehensive feature vector from d c dimensional space is mapped to the feature dimension d of the hidden layer h。 The dimension of the bias vector b1 is Base value used to adjust the mapping results.

[0166] The feature activation function g(x) is defined as

[0167] g(x)=x / (1+|x|)

[0168] Its role is to enhance nonlinear expression capabilities while compressing the range of eigenvalues ​​to ensure the stability of data in subsequent calculations. The activation function is designed to balance computational efficiency and gradient optimization performance, and its characteristics are suitable for feature inputs with a dynamic range.

[0169] After being processed by formula (8), the generated preliminary feature vector F a Enter the risk feature transformation stage. Formula (10) is used to generate the risk feature vector F r, the core of which is through W t The multi-layer linear transformation of W and W2, as well as the joint operation of tanh nonlinear activation function and element-by-element operation ⊙, extracts the characteristic information related to pancreatic cancer risk. t is the adaptive weight generation matrix, with dimension Used to capture the preliminary feature vector F a The purpose of the element-by-element multiplication operation ⊙ is to a The nonlinear transformation results are assigned dynamic weights to further enhance the sensitivity of the features to risk assessment.

[0170] The risk feature weight matrix W2 maps the above operation results to the final risk feature space, and its dimension is where d r is the dimension of the risk feature vector, which is usually set to a reasonable range of model capacity, such as 64 or 128. The bias vector b2 is used to compensate for the offset error in the feature transformation and has the dimension Risk feature vector F r It is the representation of multimodal features that is ultimately used for risk probability calculation, and includes the comprehensive results of multi-layer feature transformation and dynamic weighting.

[0171] Formula (11) transforms the risk feature vector F r Transformed into pancreatic cancer risk probability distribution P. Among them, the dimension of the weight matrix W3 is It is used to map the high-dimensional risk feature vector to the output space, and its output corresponds to two types of probabilities: healthy and sick. The dimension of the bias vector b3 is Used to adjust the balance of the final output. The Softmax activation function normalizes the result of the linear transformation into a probability distribution, which is defined as:

[0172]

[0173] where z i is the linear transformation result of the i-th category, P1 and P2 represent the probabilities of health and disease, respectively, and P1+P2=1. This normalization process ensures that the output is statistically significant and can intuitively represent the risk of pancreatic cancer for patient Luo.

[0174] For specific choices of coefficients, such as the hidden layer dimension d h , risk characteristic dimension d r , can be tuned according to the actual application scenario and the size of the training data set. The initial recommended value is d h =128,d r=64, weight matrices W1, W2, W3 and bias vectors b1, b2, b3 are all learned through the training process. Through the above design, the risk probability calculation unit can make full use of the multimodal information in the comprehensive feature vector to achieve high-precision risk probability prediction.

[0175] The early warning trigger module 104 is connected to the risk assessment module and is used to compare the pancreatic cancer risk probability with a preset threshold, generate an early warning signal when the risk probability exceeds the preset threshold, and push the early warning signal to the medical service terminal to suggest further diagnosis.

[0176] The early warning trigger module 104 is a key component in the system responsible for converting the assessment results into specific action recommendations. Its main function is to decide whether to generate an early warning signal through dynamic comparison and intelligent analysis after the risk assessment module 103 calculates the risk probability of pancreatic cancer, and push the early warning information to relevant parties in an appropriate form, thereby achieving timely intervention for individuals at high risk of pancreatic cancer.

[0177] First, the early warning trigger module receives the pancreatic cancer risk probability value from the risk assessment module and compares the value with the risk threshold pre-set in the system. In order to improve flexibility and adaptability, the threshold is not a fixed value, but is dynamically adjusted according to the user's individual characteristics and health history data. For example, for users with a history of pancreatic-related diseases, the module will appropriately lower the threshold to increase sensitivity, while for users with no relevant medical history but showing clinical symptoms, the module will assign dynamic weights according to the severity of the symptoms, thereby adjusting the trigger criteria.

[0178] When the risk probability exceeds the threshold, the early warning trigger module generates a corresponding early warning signal. The module designs a hierarchical early warning mechanism, which divides the early warning signals into different priorities based on the specific numerical value of the risk probability and the comprehensive analysis of the contextual data. Low-priority early warnings are usually used to remind users to pay attention to their health status and recommend regular check-ups, such as rechecking every six months. Medium-priority early warning signals are pushed to the user's attending physician or health management team, and it is recommended to arrange further blood tests, imaging examinations or other clinical evaluations. High-priority early warning signals are generated when the user's risk probability is extremely high or multiple indicators are abnormal, and are directly pushed to the medical institution's emergency notification system, accompanied by detailed user health data and risk analysis reports, so that medical institutions can start relevant diagnosis and treatment processes as soon as possible.

[0179] In order to ensure that the early warning signals can be processed by relevant parties in a timely manner, the module supports multi-channel signal push function. Low-priority signals are usually notified to the user through mobile applications or text messages; medium-priority signals can be sent to the user's attending physician via email or medical information system; high-priority signals are directly sent to medical institutions through a dedicated medical network in an encrypted manner. In addition, the module can also generate various forms of early warning content, including graphical risk change trends, text-described health advice, and system-generated diagnosis and treatment priority recommendations, which are convenient for the recipient to quickly understand and respond.

[0180] The warning trigger module also designs a feedback loop that allows users or doctors to mark and evaluate the effectiveness of warning signals. These feedback data will be recorded and transmitted back to the risk assessment module and the system's central database to optimize model parameters and adjust future warning strategies. For example, if doctors find that certain types of low-priority warnings appear frequently but have no practical significance, the system will automatically increase the threshold for triggering these warnings to reduce invalid alarms.

[0181] In addition, the warning trigger module has context-awareness, which means it can further refine warning decisions based on the user's real-time health data and environmental information. For example, when the system detects that the user's current lifestyle (such as diet, sleep, and exercise) may cause short-term fluctuations in risk probability, the module will temporarily suspend the generation of warning signals and observe subsequent data changes to avoid unnecessary intervention. This dynamic context-based adjustment mechanism greatly improves the accuracy and practicality of warnings.

[0182] In order to ensure the legitimacy and data security of the early warning signals, the module protects the pushed data through encryption technology and sets the access scope according to the permissions of different recipients. For example, the signal pushed to the user may only contain the risk level and simple health advice, while for medical institutions it may contain more detailed health history and specific diagnostic support data.

[0183] Through the implementation of the above functions, the early warning trigger module 104 not only provides an intelligent and personalized early warning method, but also achieves higher accuracy and efficiency through dynamic adjustment and context analysis, providing strong support for the early detection and intervention of pancreatic cancer.

[0184] Furthermore, the early warning trigger module is specifically used for:

[0185] When the risk probability is in the mild risk range, a low-priority warning signal is generated and pushed to the user terminal. It is recommended to monitor and record health data regularly;

[0186] When the risk probability is in the medium risk range, a medium priority warning signal is generated and pushed to the user's attending physician's terminal, suggesting that imaging examinations and blood tests be arranged;

[0187] When the risk probability is in a high-risk range, a high-priority warning signal is generated and pushed to the emergency notification system of the medical institution, along with the user's key health data, and it is recommended to immediately start a comprehensive diagnostic process.

[0188] The early warning trigger module in this embodiment has a highly intelligent design. Its function is not limited to generating a single risk signal, but through a multi-level response mechanism, combined with the user's risk probability interval, it dynamically generates differentiated early warning signals and pushes them to relevant recipients in an appropriate form. Such a design significantly improves the response efficiency and accuracy of the system, and can better support personalized management and timely intervention of early pancreatic cancer risks.

[0189] When the risk probability is in the mild risk interval, the early warning trigger module recognizes that the user's health status is still within a relatively safe range, but has certain potential risks. To avoid excessive intervention, the module generates a low-priority early warning signal and pushes the signal directly to the user terminal, such as a smartphone or a health management platform. In the pushed content, the module will provide suggestions based on current health data, such as regularly monitoring specific indicators (such as blood sugar levels or weight) or recording daily living habits (such as diet and sleep). In addition, the module generates a graphical risk change trend so that users can intuitively understand the dynamic changes in their health status.

[0190] When the risk probability enters the moderate risk interval, the early warning trigger module detects that the user's health status may show signs of early lesions, which require further medical examinations to confirm. In this scenario, the module generates a medium-priority early warning signal and pushes the signal to the user's attending physician's terminal, along with a detailed summary of health data, such as recent laboratory test results, imaging test reports, and symptom descriptions. The module will also recommend that doctors arrange targeted imaging examinations (such as CT or MRI) and blood tests (such as review of CA19-9 and CEA levels) to assess whether the user's health risks are likely to develop further. In order to improve diagnostic efficiency, the module will provide doctors with a preliminary analysis report based on the risk assessment model, including possible diagnostic directions and detection priority recommendations.

[0191] When the risk probability reaches the high risk range, the early warning trigger module recognizes that the user's health status may be close to critical, and comprehensive diagnosis and treatment measures need to be taken immediately. At this time, the module generates a high-priority early warning signal and pushes the signal to the emergency notification system of the medical institution through encrypted communication, along with the user's key health data. These data include the user's latest test results, the detailed calculation process of the risk probability value, dynamic trend analysis, and diagnostic assistance suggestions generated by the module. To ensure that medical institutions can respond quickly, the module will also automatically match medical resources in the user's area, such as recommending the nearest professional pancreatic cancer diagnosis and treatment center or expert team, and attaching relevant contact information and appointment procedures in the notification. Such high-priority signals are designed to prompt medical institutions to immediately start a comprehensive diagnostic process, such as arranging a pancreatic tissue biopsy or a higher-resolution imaging examination.

[0192] The early warning trigger module not only implements multi-level response through risk probability values, but also dynamically optimizes the generation and push strategy of early warning signals in combination with the user's personalized health data and medical resource conditions. For example, in the mild risk range, if the module detects that a key indicator of the user (such as the jaundice index) has a significant fluctuation in the short term, it will temporarily increase the warning priority and upgrade the mild signal to a moderate signal. Similarly, in high-risk situations, if the module finds that there is a relatively stable long-term trend in the user's historical health data, it will prompt medical institutions to avoid misdiagnosis or excessive examinations during emergency treatment.

[0193] Through this design, the early warning trigger module significantly improves the accuracy, efficiency and user experience of the system. Its multi-level response mechanism can flexibly adapt to the health management needs of different risk levels, which is a clear technical difference from the early warning mode triggered by a single threshold in the existing technology. At the same time, the module can also provide targeted health advice or auxiliary diagnosis and treatment information at each level of response, so that users, doctors and medical institutions can all benefit from it.

[0194] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. An early monitoring system for pancreatic cancer based on big data, characterized in that: include: A data acquisition module is used to obtain the user's laboratory test data, imaging test data and clinical symptom data, wherein the laboratory test data include serum CA19-9 level, serum CEA level and blood sugar level, the imaging test data include pancreatic parenchyma density value and pancreatic duct dilatation degree data, and the clinical symptom data include abdominal pain degree, weight change and jaundice index; A feature extraction module, connected to the data acquisition module, for extracting feature parameters from the acquired data and constructing a feature vector, wherein the feature vector includes a biochemical index feature sub-vector, an image feature sub-vector and a symptom feature sub-vector; A risk assessment module, connected to the feature extraction module, for obtaining the risk probability of pancreatic cancer based on the feature vector using a pre-trained deep learning model; The early warning trigger module is connected to the risk assessment module and is used to compare the pancreatic cancer risk probability with a preset threshold, generate an early warning signal when the risk probability exceeds the preset threshold, and push the early warning signal to the medical service terminal to suggest further diagnosis.

2. The pancreatic cancer early monitoring system based on big data according to claim 1 is characterized in that: The data acquisition module is specifically used for: Time series features were generated based on the change rate of serum CA19-9 levels in laboratory examination data collected at consecutive time points and the dynamic changes in the degree of pancreatic duct dilatation in imaging examination data.

3. The pancreatic cancer early monitoring system based on big data according to claim 1 is characterized in that: The feature extraction module is specifically used for: Receive continuous time point data of serum CA19-9 level, serum CEA level and blood glucose level, and normalize the data of each indicator to eliminate the influence of different dimensions; Based on the normalized data of multiple time points, calculate the change trend and fluctuation range of each indicator, and identify abnormal feature points that exceed the preset threshold; The distribution pattern of the abnormal feature points is compared with historical data to generate a set of time series feature values ​​as the basis of the biochemical indicator feature sub-vector.

4. The pancreatic cancer early monitoring system based on big data according to claim 1 is characterized in that: The feature extraction module is specifically used for: Receiving data on the density value of the pancreatic parenchyma and the degree of dilatation of the pancreatic duct, and preprocessing the image data, including denoising and segmenting the target area; Extract pancreatic edge structure and pancreatic duct morphological features from the segmented image region, including the length, diameter and curvature of the pancreatic duct; The extracted morphological features are combined with the pancreatic density value to construct the image feature sub-vector.

5. The pancreatic cancer early monitoring system based on big data according to claim 1 is characterized in that: The feature extraction module is specifically used for: Receive the user's self-reported data on abdominal pain, weight change, and jaundice index, and classify the data to distinguish between acute and chronic symptoms; For acute symptoms, assess their frequency and duration; For chronic symptoms, analyze their cumulative change trends; The symptom severity scores are compared with the user's previous health records to generate a symptom feature subvector.

6. The pancreatic cancer early monitoring system based on big data according to claim 1 is characterized in that: The early warning trigger module is specifically used for: When the risk probability is in the mild risk range, a low-priority warning signal is generated and pushed to the user terminal. It is recommended to monitor and record health data regularly; When the risk probability is in the medium risk range, a medium priority warning signal is generated and pushed to the user's attending physician's terminal, suggesting that imaging examinations and blood tests be arranged; When the risk probability is in a high-risk range, a high-priority warning signal is generated and pushed to the emergency notification system of the medical institution, along with the user's key health data, and it is recommended to immediately start a comprehensive diagnosis process.

7. The pancreatic cancer early monitoring system based on big data according to claim 1 is characterized in that: The deep learning model includes a feature fusion network layer, a temporal feature extraction layer and a risk prediction layer; The feature fusion network layer is used to receive and process input features of different modalities, and includes a first feature processing unit, which is used to reduce the dimension and extract features of the biochemical index feature subvectors through three fully connected layers, and use a nonlinear activation function to optimize the expression of features to obtain biochemical feature representation; A second feature processing unit is used to extract the spatial feature relationship of the image data by using a convolution operation, and generate an image feature representation through a fully connected layer; The third feature processing unit is used to perform feature mapping on the symptom feature sub-vector to obtain the symptom feature representation; the dynamic weighted fusion unit is used to receive the biochemical feature representation, the image feature representation and the symptom feature representation; and based on the received data, calculate the feature weight using the attention mechanism to generate a fused feature vector; The time series feature extraction layer is used to analyze the time evolution law of the fused feature vector, including a time series data cache unit, which is used to maintain a set of historical fused feature vectors and construct a time series feature matrix; A gated recurrent unit is used to analyze the time evolution law of the time series feature matrix and generate a hidden state vector; a multi-head attention mechanism is used to enhance the features of the hidden state vector and generate an enhanced time series feature vector; The risk prediction layer is used to generate the final risk assessment result, including a feature aggregation unit, which is used to receive the fused feature vector and the enhanced time series feature vector, realize feature merging through a skip connection structure, and generate a comprehensive feature vector by using a dimensionality reduction operation; The risk probability calculation unit is used to perform feature mapping on the comprehensive feature vector through multiple fully connected layers and output the pancreatic cancer risk probability distribution using the Softmax function.

8. The pancreatic cancer early monitoring system based on big data according to claim 7 is characterized in that: The dynamic weighted fusion unit is specifically used for: Receiving biochemical signatures b , image feature representation F i , Symptoms and characteristics s , and map them to the unified feature space through nonlinear functions through the following formula (1): T k =tanh(W k ·F k +b k ),k∈{b,i,s} (1) Among them, W k is the feature projection matrix; b k is the bias vector; T k Indicates the biochemical signature received F b , image feature representation F i , Symptoms and characteristics s The unified feature representation of any feature representation in , where T b is a unified feature representation of biochemical features, T i is the unified feature representation of image features, T s It is a unified feature representation of the symptom characteristics; According to the following formula (2), the fusion feature vector F is calculated c : Where λ is the adjustment factor; where angle φ k Calculate according to the following formula (3): Among them, T m is the feature vector, T m =T b +T i +T s .

9. The pancreatic cancer early monitoring system based on big data according to claim 7 is characterized in that: The multi-head attention mechanism in the temporal feature extraction layer achieves feature enhancement of the hidden state vector through the following steps: The hidden state vector output by the gated loop unit Decompose into multiple sub-vector groups according to the time step T and feature dimension d, corresponding to each attention head k, where d k =d / m represents the feature dimension of attention head k, and m is the number of attention heads; For each attention head k, the query matrix Q is generated according to the following formula (4): k , key matrix K k Sum value matrix V k : in, is the weight matrix, The attention weight matrix A is calculated according to the following formula (5): k : in, represents the average value of the key matrix, which is used to reflect the characteristic deviation between time steps; γ is the adjustment coefficient; According to the following formula (6), the output O of attention head k is generated k : Using attention weight A k Pair matrix V k Perform weighted summation to generate the output O of each head k : About k =A k ·In k (6) Among them, V k is the value matrix; According to the following formula (7), the outputs of all attention heads (O1, O2, ..., O m ) is combined into an enhanced time series feature vector F through concatenation operation e : F e =Concat(O1,O2,…,O m )·W o (7) in, is the linear transformation matrix used to map to the final output dimension.

10. The pancreatic cancer early monitoring system based on big data according to claim 7, characterized in that: The risk probability calculation unit in the risk prediction layer is specifically used for: Receive the comprehensive feature vector F output by the feature aggregation unit c ,in, d c represents the dimension of the comprehensive feature vector; According to the following formula (8), the comprehensive feature vector F is activated by the feature activation function g(x) c Perform nonlinear mapping to generate the feature vector F after preliminary processing a : F a =g(W1·F c +b1) (8) Among them, W1 is the feature mapping matrix, d h is the feature dimension of the hidden layer; b1 is the bias vector, g(x) is the activation function, which is implemented using the following formula (9): According to the following formula (10), the risk feature vector F is generated r : F r =W2·(F a ⊙tanh(W t ·F a ))+b2 (10) Among them, W2 is the risk characteristic weight matrix, d r is the dimension of the risk feature vector; W t Generate the matrix for adaptive weights, ⊙ represents element-by-element multiplication; b2 is the bias vector, According to the following formula (11), the pancreatic cancer risk probability distribution P is generated: <h2 style=";text-align:left;direction:ltr">P = Softmax(W3·F<h2 style=";text-align:left;direction:ltr"> r <h2 style=";text-align:left;direction:ltr"> +b3) (11) Among them, W3 is the weight matrix, b3 is the bias vector, Softmax is the activation function.

Images